Aidaptive x Deuna › Performance Metrics
DNN Model  ·  Volaris  ·  Processor Selection
TF DNN Model

Volaris Processor Selection
Performance Dashboard

Real-time approval rate monitoring across WORLDPAY, ELAVON, MIT_BULK, and AMEX processors. Data sourced from Snowflake PAYMENT_ML.

Total Attempts (30d)
881,979
Weighted Approval
80.0%
Processors
4
Analysis Window
30 Days
2026-03-01 to 2026-03-31
◆Processor Overview
Approval rates and volume across all four processors. May 17 - Jun 18, 2026 (33 days)
AMEX
90.71%
69,618 attempts
+0.35 pp vs prev week
MIT_BULK
81.09%
473,635 attempts
-1.65 pp vs prev week
ELAVON
87.57%
35,748 attempts
-0.62 pp vs prev week
WORLDPAY
74.86%
302,978 attempts
+0.73 pp vs prev week
Volume Distribution | Mar 8 - Apr 8
Approval Rate Comparison | Mar 8 - Apr 8
◎Recommendations
Actionable items to improve model performance.
Priority Action Impact Effort
P0 Stop WORLDPAY fallback for MXN domestic routes ~22K attempts/week rescued from 0-11% approval Model retrain or rule override
P0 Fix or remove Y4 ADYEN ECOMM MXN route 491 attempts at 15.89% approval Routing rule change
P1 Investigate WORLDPAY credit card BIN failures Improve 58.51% credit card approval BIN-level analysis
P1 Retrain model with processor-currency affinity Prevent WORLDPAY+MXN toxic combinations Feature engineering + retrain
P2 Expand 3DS enrollment for high-value transactions 3DS routes achieve 99-100% approval Integration work
P2 Document ELAVON Mar 27 configuration change Understand +19pp spike for reproducibility Investigation
◍Key Findings
Critical observations from the past month's model performance. May 17 - Jun 18, 2026
1. WORLDPAY MXN Domestic Fallback Failure: When WORLDPAY receives overflow from MIT_BULK-primary MXN domestic strategies, approval rates collapse to 0-11%. This affects ~22K attempts/week and is the single largest source of failed transactions. The model must stop routing MXN domestic to WORLDPAY as fallback.
2. Y4 ADYEN ECOMM MXN: Only 15.89% approval rate on 491 attempts routed through WORLDPAY. The ADYEN integration via WORLDPAY is fundamentally broken for MXN.
3. 3DS Routes Near-Perfect: All 3DS-authenticated strategies show 99-100% approval rates regardless of processor. Expanding 3DS enrollment could significantly boost overall approval.
4. ELAVON Improvement: +4.07 pp week-over-week improvement, with a notable spike on Mar 27 (69% to 88%). Likely a configuration change. If intentional, this is a positive signal.
5. WORLDPAY Credit Card Gap: Credit card approval at 58.51% vs debit at 66.66%. This 8pp gap needs BIN-level investigation to identify problematic card issuers.
6. Retry Approval Rates Near Zero: Retries across all processors show 0.00%–5.46% approval (vs 90–96% on first attempts). WORLDPAY: 139,843 retries at 5.46%. MIT_BULK: 101,019 retries at 0.78%. ELAVON: 16,762 retries at 0.41%. AMEX: 3,332 retries at 0.00%. This ~90pp gap suggests retries are being sent to processors that will almost certainly decline again.
7. Massive Retry Volume Waste: 260,956 retry attempts across all processors with near-zero success. This represents significant wasted processing cost, potential card network penalties, and degraded user experience. Retry logic needs to consider decline reason codes and route to different processors.
◈Daily Approval Rate Trends
Day-by-day approval rate for each processor. May 17 - Jun 18, 2026 (33 days)
ELAVON spike: Approval rate jumped from 69% (Mon-Tue) to 88% (Thu-Sat), then settled at 82-83%. Investigate possible configuration change on March 27.
WORLDPAY uptrend: Gradual improvement from 70.28% to 74.58% over the week (+4.3 pp).
◉Processor Detail
Detailed breakdown of attempts, successes, errors, and approval rates. May 17 - Jun 18, 2026
Processor Total Attempts Successes Errors Approval Rate Volume Share WoW Delta
MIT_BULK 125,906 102,876 23,030 81.71% 56.2% -0.21 pp
WORLDPAY 74,679 54,441 20,238 72.90% 33.4% -1.00 pp
AMEX 15,476 14,147 1,329 91.41% 6.9% -0.35 pp
ELAVON 7,862 6,251 1,611 79.51% 3.5% +4.07 pp
◊Strategy Analysis
Top routing strategies by volume and their approval rates. May 17 - Jun 18, 2026
Top 15 Strategies (by volume) | Mar 8 - Apr 8
Strategy Total Successes Approval Rate
Y4 CREDIT WEB CEBUR MXN91,95860,65665.96%
Y4 ECOMM USD30,77624,91180.94%
Y4 NON-CREDIT WEB MXN29,29517,30959.09%
Y4 ATO/CCL MXN16,77811,00065.56%
Y4 CREDIT WEB CEBUR MXN - DIGITALAPI14,96110,07367.33%
Y4 WP ECOMM BAN-SAN MXN14,7569,28262.90%
Y4 AMEX WEB MXN9,9989,36793.69%
Y4 ATO/CCL USD7,0286,15187.52%
3DS WP Y4 ECOMM USD X BIN5,1465,14399.94%
Y4 ECOMM USD ROW4,2393,16774.71%
N3 WP ECOMM USD2,3561,92681.75%
Y4 AMEX WEB USD2,2402,08392.99%
Y4 WP MOTO BAN-SAN MXN1,8191,30671.80%
Q6 ECOMM USD ROW1,19191576.83%
Y4 MOTO USD ROW1,06479174.34%
Critical: The highest-volume strategy "Y4 CREDIT WEB CEBUR MXN" (92K attempts) runs at only 65.96%. Within this strategy, MIT_BULK achieves 76.23% but WORLDPAY fallback collapses to 8.68%.
Multi-Processor Split Routes (WORLDPAY fallback problem) | Mar 8 - Apr 8
Strategy MIT_BULK Rate MIT_BULK Vol WORLDPAY Rate WP Vol Delta
Y4 CREDIT WEB CEBUR MXN 76.23% 77,979 8.68% 13,979 -67.55 pp
Y4 NON-CREDIT WEB MXN 69.78% 23,941 11.24% 5,354 -58.54 pp
Y4 ATO/CCL MXN 76.27% 14,026 11.01% 2,752 -65.26 pp
Y4 CREDIT WEB CEBUR MXN - DIGITALAPI 78.07% 12,903 0.00% 2,058 -78.07 pp
◇Decline Analysis
Hard decline vs soft decline breakdown by processor. May 17 - Jun 18, 2026
Decline Type Distribution | Mar 8 - Apr 8
Decline Breakdown Table | Mar 8 - Apr 8
Processor Approved Soft Decline Hard Decline
AMEX100.00%0.00%0.00%
MIT_BULK100.00%0.00%0.00%
ELAVON100.00%0.00%0.00%
WORLDPAY66.55%28.26%5.19%
WORLDPAY soft decline rate (28.26%) is nearly double MIT_BULK's (15.96%). Soft declines are often retryable. WORLDPAY also has the highest hard decline rate at 5.19%.
○Card Type Breakdown
Approval rates segmented by card type (credit, debit, prepaid) per processor. May 17 - Jun 18, 2026
WORLDPAY credit cards: Only 58.51% approval rate vs 66.66% for debit -- an 8pp gap. ELAVON shows a similar pattern (62.88% credit vs 76.90% debit).
▶Retry Metrics
Retry vs first-attempt approval rates across processors and models. May 17 - Jun 18, 2026
Why retries matter: Retry attempts represent transactions that failed on the first processor and were re-routed. Comparing first-attempt vs retry approval rates reveals which processors recover well on retries and which models route retries most effectively.
First Attempt Rate
--
Across all processors
Retry Rate
--
Across all processors
Retry Recovery Lift
--
First attempt - retry delta
Retry Volume %
--
% of total attempts that are retries
First Attempt vs Retry — by Processor
Retry Volume Share — by Processor
Retry Breakdown by Processor
Processor 1st Attempt Vol 1st Attempt Rate Retry Vol Retry Rate Delta
AMEX69,39590.65%00.00%-90.65 pp
ELAVON39,15683.87%00.00%-83.87 pp
MIT_BULK560,56967.94%00.00%-67.94 pp
WORLDPAY369,30661.61%00.00%-61.61 pp
Approval Rate by Attempt Number
Processor Attempt #1 Attempt #2 Attempt #3 Attempt #4 Attempt #5
AMEX89.96% (60,323)89.92% (3,482)77.46% (71)——
ELAVON88.42% (29,298)65.37% (2,761)43.06% (72)——
MIT_BULK85.64% (427,452)51.14% (29,442)0.00% (12,995)0.00% (58)—
WORLDPAY13.21% (64,909)94.65% (200,911)69.42% (27,184)99.19% (743)100.00% (2,492)
WORLDPAY retries actually improve: Attempt #1 at 68.54%, but attempts #2-3 jump to 81-82%, and #4-5 hit 99%+. This suggests WORLDPAY retry logic is effective — failed first attempts get recovered on subsequent tries. In contrast, ELAVON and MIT_BULK retries collapse to near 0% on attempt #2+.
Retry Performance by Model / Strategy
Retry Breakdown by Model × Processor
Model / Strategy Processor 1st Attempt Vol 1st Attempt Rate Retry Vol Retry Rate Delta
Note: applied_strategy contains routing rules, not Athia model names. Needs migration to athia_experiment_info:experiment_group for proper model attribution.
Critical finding: Retry approval rates are near zero across all processors (0.00%–5.46%). First-attempt rates are 90–96%. This ~90pp gap means retries are almost never succeeding. Possible causes: (1) retries sent to same processor that declined, (2) hard declines being retried, (3) retry logic not considering decline reason codes.
WORLDPAY has the highest retry volume (139,843 attempts) with only 5.46% approval. MIT_BULK retries 101,019 attempts at 0.78%. These represent significant wasted processing cost and potential card network penalties.
◌Week-over-Week Comparison
Comparing this week (Mar 24-31) vs previous week (Mar 17-23). WoW snapshot as of Mar 31, 2026
Processor Prev Week Rate Prev Week Vol This Week Rate This Week Vol Delta
AMEX90.44%14,43190.79%17,711 +0.35 pp
ELAVON87.80%8,64287.18%9,278 -0.62 pp
MIT_BULK81.69%84,91280.04%97,308 -1.65 pp
WORLDPAY76.42%58,75277.15%98,986 +0.73 pp
★Volaris Smart Router Analysis
Deep-dive into routing strategy performance, decline reasons, and error categories for Volaris. 30-day window
Note: athia_experiment_info:experiment_group is not yet populated in transactions. Strategy names below are Deuna routing rules. Once experiment groups are enabled, this section will show per-model (TF Per-Processor vs XGBoost LTR) comparisons.
Total Volaris Attempts
1,037,928
30-day window
Top Strategy
99.9%
3DS WP Y4 ECOMM USD X BIN [WORLDPAY]
Worst Strategy
1.1%
Y4 ADYEN ECOMM BBVA WEB MXN [WORLDPAY]
Routing Rules Active
57
Distinct strategies
Top Routing Strategies — Performance by Processor
Strategy Performance Table
StrategyProcessorVolumeApproval RateErrors
Y4 CREDIT WEB CEBUR MXNMIT_BULK310,61873.66%81,825
Y4 NON-CREDIT WEB MXNMIT_BULK112,43662.30%42,384
Y4 ECOMM USDWORLDPAY100,66485.34%14,760
Y4 CREDIT WEB CEBUR MXN - DIGITALAPIMIT_BULK59,37066.05%20,155
Y4 AMEX WEB MXN AMEX49,41292.88%3,517
Y4 CREDIT WEB CEBUR MXNWORLDPAY47,74622.88%36,820
Y4 ATO/CCL MXNMIT_BULK47,34076.61%11,074
Y4 WP ECOMM BAN-SAN-BBVA-BNMX MXNWORLDPAY44,33968.72%13,870
Y4 WP ECOMM BAN-SAN MXNWORLDPAY43,98876.54%10,318
Y4 ATO/CCL USDELAVON27,70289.51%2,906
Y4 NON-CREDIT WEB MXNWORLDPAY26,85417.74%22,090
Y4 ECOMM USD ROWWORLDPAY17,01177.36%3,851
Y4 CREDIT WEB CEBUR MXN - DIGITALAPIWORLDPAY15,35514.12%13,187
Y4 WP ECOMM BAN-SAN-BBVA-BNMX MXNMIT_BULK11,91718.85%9,671
Y4 ATO/CCL MXNWORLDPAY10,35426.86%7,573
Y4 AMEX WEB USD AMEX9,80492.52%733
Y4 ECOMM USDELAVON9,27069.67%2,812
3DS WP Y4 ECOMM USD X BINWORLDPAY8,81699.92%7
N3 WP ECOMM USDWORLDPAY8,74381.15%1,648
Y4 WP ECOMM BAN-SAN MXNMIT_BULK8,41512.01%7,404
Y4 MOTO USD ROWWORLDPAY4,99275.06%1,245
Y4 ADYEN WEB BBVA DEBITO MXNWORLDPAY4,5545.88%4,286
Y4 DEBIT WEB MXN - DIGITALAPIWORLDPAY4,53020.44%3,604
Y4 WP MOTO BAN-SAN MXNWORLDPAY4,47880.01%895
Y4 WP MOTO BAN-SAN-BBVA-BNMX MXNWORLDPAY4,12673.80%1,081
Error Categories — by Processor
Decline Type Mix — by Processor
Error Category Breakdown
ProcessorError CategoryCount% of Errors
AMEXNA_no_identificado3,68457.37%
AMEXbank_rejected2,62640.90%
AMEXinsufficient_amount1111.73%
ELAVONbank_rejected2,26636.12%
ELAVONNA_no_identificado1,02116.27%
ELAVONinsufficient_amount80612.85%
ELAVONfraud_psp78112.45%
ELAVONuser_settings69311.05%
ELAVONpsp_int_settings3235.15%
ELAVONmax_attemps_quantity1812.88%
ELAVONfraud_bank1402.23%
ELAVONdeuna_settings631.00%
MIT_BULKbank_rejected125,43769.82%
MIT_BULKinsufficient_amount15,2058.46%
MIT_BULKuser_settings12,9477.21%
MIT_BULKpsp_int_settings9,7555.43%
MIT_BULKNA_no_identificado7,8254.36%
MIT_BULKdeuna_settings3,6102.01%
MIT_BULKfraud_psp2,0581.15%
MIT_BULKfraud_bank1,4470.81%
MIT_BULKmax_attemps_quantity1,0700.60%
MIT_BULKpsp_ext_settings3150.18%
WORLDPAYbank_rejected97,52968.79%
WORLDPAYuser_settings8,7806.19%
WORLDPAYpsp_int_settings7,6615.40%
WORLDPAYinsufficient_amount7,5885.35%
WORLDPAYNA_no_identificado6,0514.27%
WORLDPAYfraud_bank4,2693.01%
WORLDPAYfraud_psp4,1882.95%
WORLDPAYmax_attemps_quantity2,8382.00%
WORLDPAYpsp_ext_settings1,5961.13%
WORLDPAYdeuna_settings1,2780.90%
Toxic Routing Combinations (Approval < 15%)
StrategyProcessorVolumeApproval RateAction
Y4 CREDIT WEB CEBUR MXN - DIGITALAPIWORLDPAY15,35514.12%Block or reroute
Y4 WP ECOMM BAN-SAN MXNMIT_BULK8,41512.01%Block or reroute
Y4 ADYEN WEB BBVA DEBITO MXNWORLDPAY4,5545.88%Block or reroute
Y4 DEBIT WEB MXN - DIGITALAPIMIT_BULK2,3463.41%Block or reroute
Y4 ADYEN ECOMM BBVA WEB MXNWORLDPAY1,8291.09%Block or reroute
Y4 ADYEN WEB BBVA DEBITO MXNMIT_BULK1,6907.10%Block or reroute
Y4 ADYEN ECOMM BBVA WEB MXNMIT_BULK9296.46%Block or reroute
Toxic combinations are strategy+processor pairs with <15% approval on significant volume. These represent wasted attempts that should be blocked by the model or routing rules.
Daily $ Lift — DNN vs Heuristic (Volaris, 30-day)

Daily transaction value lift of DNN over heuristic in MXN. Positive = DNN winning. Lift = (DNN approval rate − Heuristic approval rate) × DNN volume × avg transaction amount.

30-Day Cumulative Lift (MXN)
—
— of — days DNN winning
Avg Daily Lift (MXN)
—
per active day
Avg DNN Approval Rate
—
Heuristic: —
⚙Model Performance — DNN vs Heuristic
Comparison of ML model types across accuracy, approval rate, latency, and volume. 7-day window
Prediction Accuracy
DNN Accuracy
73.51%
11,995 predictions
Heuristic Accuracy
37.50%
12,226 predictions
DNN vs Heuristic
+57.05 pp
accuracy difference
Daily Prediction Accuracy by Model
Daily Approval Rate by Model — Per Processor
MIT_BULK
WORLDPAY
AMEX
ELAVON
Accuracy by Model × Processor (7-day)
ModelProcessorTotalCorrectAccuracy
DNNELAVON14800.00%
HeuristicELAVON19200.00%
DNNMIT_BULK10,6588,81882.74%
HeuristicMIT_BULK5,9444,58577.14%
DNNWORLDPAY1,18900.00%
HeuristicWORLDPAY6,09000.00%
Approval Rate
Daily Approval Rate by Model (processed / total)
Daily Decline Rate by Model (denied / total)
Daily Approval Rate — DNN vs Amount-Matched Sample (14-day)

Non-DNN transactions sampled to match DNN volume per amount bin each day — removes transaction-mix confound vs raw heuristic comparison.

Daily Approval Rate by Model Version (14-day)
Approval Rate by Model Version (14-day)
VersionTypeTotalApprovedApproval Rate
dnn:v43:tf_per_processorDNN18815381.38%
dnn:v47:tf_per_processorDNN78762279.03%
dnn:v48:tf_per_processorDNN5,2764,04176.59%
dnn:v54:tf_per_processorDNN12,5369,27473.98%
dnn:v56:tf_per_processorDNN65145670.05%
heuristic-v0Heuristic20,40715,81777.51%
Approval Rate by Model × Processor (7-day)
ModelProcessorTotalApprovedApproval Rate
DNNELAVON14800.00%
HeuristicELAVON19200.00%
DNNMIT_BULK10,6588,83082.85%
HeuristicMIT_BULK5,9444,59377.27%
DNNWORLDPAY1,189171.43%
HeuristicWORLDPAY6,0904,55074.71%
Volume & Latency
DNN p95 Latency
16ms
avg 6.2ms
Heuristic p95 Latency
3ms
avg 1.4ms
Retry Performance
Daily Approval Rate by Routing Rule — Retries Only
Daily Retry Volume by Routing Rule (Y4 / N3 / Q6)
Daily Approval Rate by Routing Rule — Retries Only (success / total)
Approval Rate by Card Type — DNN vs Heuristic (30 days)
Volume and approval rate per card type (credit / debit / prepaid) across control and treatment groups.
Daily Retry Rate — DNN vs Heuristic (all processors combined)
% of orders that needed at least one retry attempt. Lower = better first-attempt routing.
Daily Retry Rate by Processor — DNN vs Heuristic
Per-processor breakdown. Shows whether DNN's processor selection reduces retries on each PSP.
⚖Model Comparison — Control vs Treatment
Athia feedback accuracy: control (baseline) vs treatment (Athia model) vs overall routing approval. 30-day window
Control (Baseline)
78.05%
105,275 predictions
Treatment (Athia)
66.78%
57,652 predictions
Overall Routing
79.87%
3,119,019 attempts
Treatment vs Control
-11.27 pp
Positive = treatment outperforms
Daily Accuracy Trend — Control vs Treatment vs Routing

Bars show daily transaction volume (right axis); lines show accuracy % (left axis).

Model Comparison Summary
Group Event Type Total Correct Accuracy Date Range
controlfeedback105,27582,16278.05%2026-03-05 02:54:32 to 2026-06-18 04:22:50
routingrouting3,119,0192,491,05479.87%2026-03-05 02:54:33.383000 to 2026-06-18 04:11:17.746000
treatmentfeedback57,65238,50266.78%2026-03-31 03:49:42 to 2026-06-18 04:22:35
Daily Breakdown ▼
Date Group Volume Correct Accuracy
2026-05-17control37029579.73%
2026-05-17routing17,76414,34780.76%
2026-05-17treatment39030277.44%
2026-05-18control61851282.85%
2026-05-18routing27,09422,32482.39%
2026-05-18treatment44028965.68%
2026-05-19control71656078.21%
2026-05-19routing32,21225,45779.03%
2026-05-19treatment48815531.76%
2026-05-20control69357182.40%
2026-05-20routing31,52325,46380.78%
2026-05-20treatment50037875.60%
2026-05-21control28623080.42%
2026-05-21routing29,99123,08476.97%
2026-05-21treatment20115275.62%
2026-05-22control25221183.73%
2026-05-22routing27,64221,85979.08%
2026-05-22treatment21617781.94%
2026-05-23control52242781.80%
2026-05-23routing20,99616,43978.30%
2026-05-23treatment45733372.87%
2026-05-24control43433577.19%
2026-05-24routing16,15712,58877.91%
2026-05-24treatment31323374.44%
2026-05-25control72559181.52%
2026-05-25routing36,89229,22379.21%
2026-05-25treatment62146975.52%
2026-05-26control1,3581,13583.58%
2026-05-26routing43,93134,72679.05%
2026-05-26treatment1,05481677.42%
2026-05-27control2,0871,77485.00%
2026-05-27routing41,59732,70078.61%
2026-05-27treatment1,4121,07976.42%
2026-05-28control1,9981,63181.63%
2026-05-28routing39,57430,50577.08%
2026-05-28treatment1,25382766.00%
2026-05-29control1,8721,52181.25%
2026-05-29routing37,49729,02977.42%
2026-05-29treatment1,3781,01673.73%
2026-05-30control1,5381,25281.40%
2026-05-30routing29,45522,72377.14%
2026-05-30treatment1,06278974.29%
2026-05-31control1,2481,01181.01%
2026-05-31routing24,40218,80377.06%
2026-05-31treatment90164871.92%
2026-06-01control1,8281,18164.61%
2026-06-01routing36,36528,57478.58%
2026-06-01treatment1,4191,09076.81%
2026-06-02control2,18384338.62%
2026-06-02routing44,17434,31977.69%
2026-06-02treatment2,1621,79482.98%
2026-06-03control2,07773635.44%
2026-06-03routing41,47531,99577.14%
2026-06-03treatment2,1751,73779.86%
2026-06-04control1,28349538.58%
2026-06-04routing26,77221,47480.21%
2026-06-04treatment1,24594976.22%
2026-06-05control90133837.51%
2026-06-05routing26,13721,08480.67%
2026-06-05treatment1,16988375.53%
2026-06-06control93738741.30%
2026-06-06routing20,69916,74680.90%
2026-06-06treatment75357576.36%
2026-06-07control77631440.46%
2026-06-07routing16,85613,83382.07%
2026-06-07treatment51737071.57%
2026-06-08control1,11547142.24%
2026-06-08routing24,09719,80982.21%
2026-06-08treatment70152975.46%
2026-06-09control1,09242538.92%
2026-06-09routing28,52523,35281.87%
2026-06-09treatment90565872.71%
2026-06-10control1,22944536.21%
2026-06-10routing29,40022,96078.10%
2026-06-10treatment1,14480470.28%
2026-06-11control1,00035035.00%
2026-06-11routing26,09721,08980.81%
2026-06-11treatment95069272.84%
2026-06-12control93639341.99%
2026-06-12routing25,85121,19281.98%
2026-06-12treatment96573476.06%
2026-06-13control79529837.48%
2026-06-13routing21,13917,27781.73%
2026-06-13treatment77558675.61%
2026-06-14control76228837.80%
2026-06-14routing17,18214,01481.56%
2026-06-14treatment72755676.48%
2026-06-15control1,24948138.51%
2026-06-15routing26,44921,34980.72%
2026-06-15treatment1,20492776.99%
2026-06-16control2,32092439.83%
2026-06-16routing31,97126,01481.37%
2026-06-16treatment2,1881,66776.19%
2026-06-17control3,2161,13135.17%
2026-06-17routing36,94127,80175.26%
2026-06-17treatment3,2082,21869.14%
2026-06-18control71827538.30%
2026-06-18routing8,3506,80481.49%
2026-06-18treatment83463476.02%
How to read: "Control" is the baseline routing (no Athia model). "Treatment" is the Athia ML model. "Routing" is the overall smart routing approval rate across all transactions. A higher treatment accuracy means the Athia model is predicting processor outcomes better than baseline.
Low sample size: Treatment data starts appearing Mar 31 with only 226 predictions — as the experiment ramps up, the comparison will become more statistically significant. Current delta (-1.62 pp) is not yet reliable.
⚖Sampling Bias Audit
Volaris Multi-Horizon Forensic Audit — DNN vs Heuristic population distinguishability across 3 rolling windows (24h / 7d / 15d). TVD < 2% = indistinguishable; 2-5% = minor drift; ≥5% = divergent.
24h Verdict
—
—
7d Verdict
—
—
15d Verdict
—
—
15d 6D avg TVD
—
—
TVD per Feature × Window
Total Variation Distance between control (Heuristic) and treatment (DNN) for each pre-routing feature. Lower = more similar populations.
Group Sizes per Window
WindowControl (n)Treatment (n)TotalSplit %Balance
No data — run refresh_metrics.py to populate
TVD per Feature (7d window)
FeatureTVDStatus
No data — run refresh_metrics.py to populate
⚙Model Reference collapsed
All ML models and routing strategies currently deployed in production.
▼
Production Model Types
ModelFrameworkVersionsPurpose
processor_selector XGBoost LTR v1, v2, v2_2, v3, v4 Rank and select optimal payment processor for each transaction
volaris_smart_router TensorFlow SavedModel v1 (per-processor) Per-processor approval probability prediction (140-feature encoder)
retry_predictor XGBoost LTR v1, v2, v2_1, v2_2 Predict whether a failed transaction should be retried
installment_optimizer scikit-learn RF — Predict optimal installment options for customer
Routing Strategies
StrategyModel UsedDescriptionStatus
tf_per_processor volaris_smart_router Scores each processor independently with TF model, then ranks. Default for Volaris. Default
xgboost_ltr processor_selector Scores all processors in a single XGBoost batch using Learning-to-Rank. Active
Volaris Per-Processor TF Models
ProcessorModel FormatFeaturesEncoder
worldpayTF SavedModel140 (13 z-score + 127 OHE)volaris_tf_encoder.py
amexTF SavedModel140 (13 z-score + 127 OHE)volaris_tf_encoder.py
elavonTF SavedModel140 (13 z-score + 127 OHE)volaris_tf_encoder.py
mit_bulkTF SavedModel140 (13 z-score + 127 OHE)volaris_tf_encoder.py
uatpTF SavedModel140 (13 z-score + 127 OHE)volaris_tf_encoder.py
bbvaTF SavedModel140 (13 z-score + 127 OHE)volaris_tf_encoder.py
Infrastructure
ComponentDetail
Servingathia-model-server sidecar (Python FastAPI) alongside Go API
Model StorageS3 → EFS mount at /mnt/models
Training RepoDATA-Athena-Snowflake (Python, ClearML tracking)
Serving Repoathena-platform (Go + Python sidecar)
InfraSpacelift — terraform-athia
DashboardAthia Experiments Dashboard